Oral-3D: Reconstructing the 3D Structure of Oral Cavity from Panoramic X-ray
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Oral-3D: Reconstructing the 3D Structure of Oral Cavity from Panoramic X-ray Weinan Song * , Yuan Liang * , Jiawei Yang, Kun Wang, Lei He University of California, Los Angeles, USA Abstract arXiv:2003.08413v4 [eess.IV] 9 Jan 2021 Panoramic X-ray (PX) provides a 2D picture of the patient’s mouth in a panoramic view to help dentists observe the in- visible disease inside the gum. However, it provides limited 2D information compared with cone-beam computed tomog- raphy (CBCT), another dental imaging method that gener- ates a 3D picture of the oral cavity but with more radiation dose and a higher price. Consequently, it is of great interest to reconstruct the 3D structure from a 2D X-ray image, which can greatly explore the application of X-ray imaging in den- tal surgeries. In this paper, we propose a framework, named Oral-3D, to reconstruct the 3D oral cavity from a single PX image and prior information of the dental arch. Specifically, we first train a generative model to learn the cross-dimension Figure 1: An overview of Oral-3D. We first back-project the transformation from 2D to 3D. Then we restore the shape of panoramic image into a flattened 3D image of the oral cavity the oral cavity with a deformation module with the dental arch with a generative network, then we utilize the dental arch curve, which can be obtained simply by taking a photo of the patient’s mouth. To be noted, Oral-3D can restore both information to map the generation result into a curved plane the density of bony tissues and the curved mandible surface. to reconstruct the final 3D structure. Experimental results show that Oral-3D can efficiently and effectively reconstruct the 3D oral structure and show criti- cal information in clinical applications, e.g., tooth pulling and 39.4× radiation (Brooks 2009) and takes 3.7× of the price dental implants. To the best of our knowledge, we are the first to explore this domain transformation problem between these (Petersen et al. 2014) on average than PX. This problem is two imaging methods. especially evident for those sensitive to the radiation dose and the developing countries where people are unwilling to invest much in dental healthcare. Therefore, it is of great Introduction interest to directly reconstruct the 3D structure of the oral Extra-oral imaging techniques such as PX and CBCT are cavity from a PX image. widely used in dental offices as examination methods before However, it is of great challenge to reconstruct a 3D object the treatment. Both methods can show detailed bone infor- from a single 2D image due to the lack of spatial information mation, including the tooth, mandible, and maxilla, of the in the rendering direction. Most works rely on additional in- entire oral cavity. However, during the imaging process of formation, such as shadow or prior shape of the object, to PX, the X-ray tube moves around the patient’s head and can regularize the reconstruction result. Furthermore, this prob- only take a 2D panoramic picture. This has limited its appli- lem is more difficult for the oral cavity due to the compli- cation in the cases when the disease needs to be located. In cated shape of the mandible and detailed density informa- comparison, CBCT can reconstruct the whole 3D structure tion of the teeth. To overcome such challenges, we propose of the lower head with divergent X-rays and provide abun- a two-stage framework, named Oral-3D, to generate a high- dant information about the health condition of oral cavity. resolution 3D structure of the oral cavity by decoupling the Nevertheless, the patient needs to take more radiation dose reconstruction process of shape and density. We first train and pay a higher price during a CBCT scan. We summarize a generation model to extract detailed density information the characteristics of these two imaging methods in Table 1. from the 2D space, then restore the mandible shape with the We can see that although CBCT can provide more informa- prior knowledge of the dental arch. Although our method tion in clinical applications (Momin et al. 2009), it generates can not totally replace CBCT in the dental examination, we provide a compromise solution to obtain the 3D oral struc- * Equal Contributions ture when only the PX is available.
Table 1: A comparison of CBCT and panoramic X-ray on common dental disease. Imaging Dimension Imaging Cost Radiation Dose Diagnostic Accuracy Wisdom Tooth Implant Orthodontics Method (Petersen et al. 2014) (Brooks 2009) (Momin et al. 2009) Tooth Decay Planning CBCT 3D e 184.44 58.9-1025.4 µSv 94.8% 3 3 3 3 PX 2D e 49.29 5.5-22.0 µSv 83.3% 3 3 7 7 Our work can be summarized as a combination of a method for instance-level segmentation and identification of single-view reconstruction problem and a cross-modality teeth in the CBCT image. (Lee et al. 2018) trains a deep transformation problem, where the model should recover neural network to detect and diagnose dental caries from pe- both the shape and density information of the target object riapical radiographic images. (Imangaliyev et al. 2016) de- from a single image. We show an overview of Oral-3D in signs a classification model for red auto-fluorescence plaque Figure 1. At first, we train a generation network to learn the images to assist in detecting dental caries and gum diseases. cross-dimension transformation that can back-project the 2D (Prajapati, Nagaraj, and Mitra 2017) uses transfer learning PX image into 3D space, where the depth information of to classify three different oral diseases for X-ray images. Al- teeth and mandible can be learned automatically from the though these methods have improved oral healthcare service paired 2D and 3D images. In the second step, we register by providing intelligent assistance, the model needs to be this generated 3D image, a flattened oral structure, into a trained with annotations on large datasets, which requires curved plane to restore the original shape according to the both professional knowledge and tedious labour. Compare dental arch. This prior knowledge effectively restricts the with these works, our model helps dental healthcare without shape and location of the 3D structure and can be obtained in the supervision of labelled data, where the 3D reconstruc- many ways, such as by fitting the β function with the width tion is learned from the latent relationship between 2D and and depth of the mouth (Braun et al. 1998). To show the 3D images. effectiveness of our framework, we first compare Oral-3D with other methods on synthesized images generated from a CBCT dataset. Then we evaluate the reconstruction results Cross-Modality Transfer in Medical Imaging for some clinical cases to prove the feasibility of our method. Experimental results show that Oral-3D can reconstruct the The target of cross-modality transfer is to find a non-linear 3D oral structure with high quality from a single panoramic relationship between medical images in different modali- X-ray image and keep the density information simultane- ties. It can help reduce the extra acquisition time and ad- ously. In conclusion, we make the following contributions: ditional radiation in a medical examination or provide ad- • We are the first to explore the cross-modality transfer of ditional training samples without repetitive annotation work images in different dimensions for dental imaging by deep to augment the dataset. Most works take this as a pix-to-pix learning. In addition to restoring the 3D shape and surface problem, where the layout and the structure are consistent, of the bone structure, our model can restore the density in- but the colour distribution is changed after the transforma- formation simultaneously, which is of great help for den- tion between images in different modalities. For example, tal diagnosis. as shown in Figure 2, (Costa et al. 2017) takes the vessel tree of eyes as a condition to synthesis new images for fun- • We decouple the reconstruction process for density and dus photography. (Choi and Lee 2018) proposes a generation shape recovery by proposing a deformation module that network to produce realistic structural MR images from flor- embeds a flattened 3D image into a curved plane. This betapir PET images. However, few studies have discussed has not been addressed in previous research and can sig- the cross-modality transfer problem from a lower-dimension nificantly improve the reconstruction performance. image to a higher-dimension one, which is more challeng- • We propose an automatic method to generate paired 2D ing as the model needs to infer high-dimension information and 3D images to train and evaluate the reconstruction from the lower-dimension image. We only find two works models, where Oral-3D achieves relatively high perfor- that achieve a similar target to ours. Specifically, (Henzler mance and can show key features of some typical cases. et al. 2018) uses an encoder-decoder network to reconstruct Meanwhile, we propose a workflow to evaluate our model 3D skull volumes of 175 mammalian species from 2D cra- on a real-world dataset, which indicates the feasibility of nial X-rays, but the result is subject to too much ambiguity. clinical applications. To improve the visual quality, (Ying et al. 2019) utilizes bi- planar X-rays to extract 3D anatomical structures of body Related Work CT with adversarial training and reconstruction constraints. However, our problem is quite different from theirs as the Deep Learning for Oral Health PX image can not be synthesized only from the orthogonal Deep learning has dramatically promoted the computer as- projection over the corresponding CT. Besides, our task is sistance system for dental healthcare by automatically learn- more challenging due to the complicated structure of the oral ing feature representations from large amounts of data. For cavity, where the model is required to restore more details of example, (Cui, Li, and Wang 2019) proposes an automatic the teeth and mandible.
inator D in an adversarial way. The generator learns to out- put a fake image from a random vector to deceive the dis- criminator, while the discriminator tries to distinguish sam- pling data between real and fake images. As we aim to gen- erate the consistent 3D content from the semantic informa- tion of the panoramic X-ray image, we utilize conditional GANs (Mirza and Osindero 2014) as the generative model Figure 2: We show some examples of cross-modality trans- to learn the back-projection transformation. fer for Retina → Fundus (Costa et al. 2017) and PET → MRI (Choi and Lee 2018), where the source image and the target image usually contains consistent physiological struc- Objective Function To improve the generation quality tures although in different modalities. and guarantee the stable training process, we use LSGAN (Mao et al. 2017) as the keystone to train the generator and discriminator. The adversarial loss can be summarized as: 3D Reconstruction from 2D Image LossD =Ey (D(y) − 1)2 + Ex D(G(x))2 Recent work of 3D reconstruction from 2D images can be concluded as two categories: multi-view reconstruction and (1) LossG =Ex D(G(x)) − 1)2 , single-view reconstruction. For the first one, the method generally requires little prior knowledge about the 3D ob- where x is the PX image and y is the flattened oral structure. ject as the images taken from multiple angles can restrict the To maintain the structural consistency of the input and reconstruction shape. For example, (Kolev, Brox, and Cre- generation result, we also introduce the reconstruction loss mers 2012) computes the most probable 3D shape that gives and projection loss to improve the generation quality. These rise to the observed colour information from a series of cal- proposed loss functions can bring voxel-wise and plane-wise ibrated 2D images. (Choy et al. 2016) learns the mapping regularization to the generation network, which can be de- function from arbitrary viewpoints to a 3D occupancy grid fined as: with a 3D recurrent neural network. As a comparison, re- construction from a single-view image usually requires ad- LossR =Ex,y (y − G(x))2 ditional information, e.g., prior shape, to inference the ob- (2) LossP =Ex,y (P (y) − P (G(x)))2 , ject shape. As such, (Yang et al. 2018) proposes a unified framework trained with a small amount of pose-annotated images to reconstruct a 3D object. (Wu et al. 2018) takes the where the function P () is achieved by orthogonal projec- adversarially learned shape priors as a regularizer to penal- tions along each dimension of the generated 3D image. In ize the reconstruction model. However, the PX image can be summary, the total optimization problem can be concluded either seen as a single-view image taken by a moving cam- as: era or a concatenate image blended with multiple views. In D∗ = arg min LossD this paper, we take our problem as the first kind to decou- D ple the reconstruction process for the bone density and the G∗ = arg min λ1 · LossG + λ2 · LossR + λ3 · LossP . mandible shape. In the experiment, we also show that this G (3) can significantly promote performance over the multi-view reconstruction model both in quality and quantity. Generator During the X-ray imaging, the depth informa- Method tion can be reflected in the absorption of radiation through In this section, we introduce our framework that reconstructs the bone. Therefore it is reasonable to extract the thickness a high-resolution 3D oral cavity from a 2D PX image. We of the tooth and the mandible from a PX image. Then the choose to break this problem into two stages to recover more objective for the generator is to find a cross-dimension trans- details of the bone density. We show the structure of Oral- formation G from 2D to 3D, which can be denoted as: 3D, which consists of a back-projection module and a defor- mation module in Figure 3. The back-projection module de- 2D 3D G : IH×W → IH×W ×D , (4) velops from generative adversarial networks (GAN (Good- fellow et al. 2014)), where the generator is trained to learn where I 2D is the PX image with a size of H × W and I 3D is a back-projection transformation by exploring the depth in- the flattened 3D structure with a size of H × W × D. In this formation contained in the X-ray image. The deformation paper, we utilize 2D convolution to retrieve the latent depth module takes in the generated 3D image (Image c) from the information. The 3D information is embedded into differ- back-projection module and the dental arch (Image e) to re- ent channels of feature maps. As shown in Fig. 3, the en- store the curved shape of the mandible. coding network decreases the resolution of feature maps but increases the number of feature channels, while the decod- Back-Projection Module ing network increases the resolution to generate a 3D object. GANs have proved to be an effective model to learn latent The output voxel value is restricted to (−1, 1) with a tanh data distribution by training the generator G and the discrim- layer at the end.
Figure 3: Our framework consists of two modules to decouple the recovery of bone density and the mandible shape. The back- projection module utilizes a generation network to restore the 3D density information from the 2D space, and the deformation module transforms the flattened 3D image into a curve plane according to the prior knowledge in the dental arch. Dense Block Dense connections (Huang et al. 2017) have paired of 3D images when training the discriminator. shown compelling advantages for feature extraction in deep neural networks. This architecture is especially efficient in Deformation Module forwarding 3D information as each channel of the output With the generation of a 3D image from the back-projection has a direct connection with intermediate feature maps. In module, the deformation model maps the flattened 3D struc- the projection module, we utilize two kinds of dense blocks, ture into the curved space according to the arch curve to noted as A and B, to extract depth information from the X- output the final reconstruction object. As shown in the right ray image. As shown at the bottom of Figure 3, the dense part of Figure 3, we propose a registration algorithm that block A explores the depth information by increasing the can best restore the shape of the oral cavity and keep the channel number of feature maps. In contrast, the dense block recovered density information. We first sample the gener- B fuses feature maps from the up − samplinglayer and ated 3D image (Image c) into slices (Image f ) in the sagit- the skip-connections but maintain the number of channels tal plane, then interpolate these slices along the dental arch to forward the depth information. In the end, the number of curve (Image e). To achieve this, we sample a number of stacked features in the output is equal to the depth of the points from the curve with equal distance and embed the generated 3D image. slices into the curve. In the end, we interpolate the voxels between the neighbouring slices to output a smooth 3D im- Discriminator The discriminator has been frequently age (Image g). For computation convenience, we combine used in many generative models to improve the generation these steps together and conclude it in Algorithm 1, where quality by introducing an instance-level loss. In the back- we assume that the generated 3D image and the bone model projection module, we adopt a patch discriminator intro- has the same height of H. duced by (Isola et al. 2017) to improve the generation quality of tooth edges by learning high-frequency structures in the Experiment flattened 3D structure. We set the patch size as 70 × 70 × 70 and follow a similar structure in (Isola et al. 2017) but re- Dataset place 2D convolution with 3D. The discrimination network As grouped data of PX image, dental arch shape, and 3D ends with a Sigmoid layer to predict the probability of the oral structure of the same patient, especially in the same pe- samples belonging to the real image. To be noted, we sample riod, is hard to find, we first use synthesized data to evaluate the same number of 3D patches at the same position from the the performance. We collect 100 CBCT scans from a major
Algorithm 1 Embedding and Interpolation 1: function REGISTER(Slices, W3D , D3D , curve) 2: W, H, D ← SHAPE(Slices) 3: OralImage ← ZEROS(W3D , H, D3D ) 4: SampleP oints ← SAMPLE(curve) 5: for i = 0; i < W3D ; i + + do 6: for j = 0; j < D3D ; j + + do 7: id, dist ← DIST((i, j), SampleP oints) 8: Slice ← Slices[id, :, :] 9: Slice ← INTERPOLATE(Slice, dist) 10: OralImage[i, :, j] ← Slice 11: end for 12: end for 13: return OralImage 14: end function stomatological hospital in China and re-sample these 3D im- Figure 4: An overview of generating paired data for 3D oral ages into a size of 288 × 256 × 160. The dataset is finally structure and 2D panoramic X-ray is shown in this picture. normalized into a range of (−1, 1) and split into a ratio of We first get the MIP image from the CBCT scan to obtain 3 : 1 : 1 for training, validation, and testing. the dental arch curve (red), and boundaries of the dental area An overview of preparing the synthesized data can be seen (blue and green). Then we obtain the flattened oral structure, in Figure 4. We first obtain a 2D image in the axial plane PX image, and the 3D oral structure by re-sampling, projec- by maximum intensity projection (MIP) (Image b) over the tion, and extraction, respectively. CBCT slices (Image a). Then we obtain the dental curve with a similar method as in (Yun et al. 2019) to estimate the curve function and boundaries of the dental arch. To gener- • Overall: To combine these three metrics together, we also ate the PX image (Image d), we simulate projection with the define a score S = (P SN R/20 + Dice + SSIM )/3 to Beer-Lambert absorption-only model along the arch curve. compare the overall performance of the 3D reconstruc- This imaging process is similar to the way for a real PX ma- tion. chine, where the manufacturer usually improves the imag- ing quality by designing a trajectory of the camera to fit Comparison Models the mandible shape. Finally, we extract the 3D oral struc- ture (Image e) by removing the unrelated tissues with the To show the effectiveness and efficiency of Oral-3D, we also boundaries and generate the flattened 3D structure (Image compare our framework with other models that work on a c) by re-sampling along the arch curve. similar problem: • Residual CNN: An encoder-decoder network that has Evaluation Metrics been introduced in (Henzler et al. 2018) to reconstruct the • PSNR: Peak signal-to-noise ratio (PSNR) is often used 3D model with a single X-ray. to measure the difference between two signals. Compared • GAN: A generative model based on (Goodfellow et al. with mean squared error, PSNR can be normalized by the 2014) that takes the Res-CNN as the backbone for gener- signal range and expressed in terms of the logarithmic ator with reconstruction loss and the same discriminator decibel scale. We take this to measure the density recov- as Oral-3D. ery of our models. • Dice: In order to reflect the deformation of the reconstruc- • R2N2: We transform our task into a multi-view recon- tion, we use dice coefficient between our reconstruction struction problem to train R2N2 (Choy et al. 2016) by tak- results and the groundtruth in a volume level of the oral ing the PX image as a composition of X-ray image taken cavity. The 3D volume of the oral cavity is obtained by from three different views. setting a threshold (e.g.,−0.8 over the reconstruction re- • Auto Encoder: We remove the discriminative network in sult. Oral-3D and keep the encoder-decoder network only in • SSIM: We use the structure similarity index (SSIM) the back-projection module. (Wang et al. 2004) as the key criterion to quantify the per- formance of density recovery.SSIM considers the bright- Training ness, contrast and structure information at the same time All the experiment are trained by Adam optimizer (Kingma and can match better the subjective evaluation of humans. and Ba 2014) with a batch size of 1 for 300 epochs. The It can effectively indicate the reconstruction quality and learning rate starts at 1 × 10−3 and decreases 10 times every is widely used in other similar works, such as (Ying et al. 50 epochs. We use the validation data as the stop criterion, 2019). and all models converge after 300 epochs. For adversarial
Figure 5: We show the qualitative comparison from different views and rendering ways in this picture. We can see that our method generates the best results with more detailed density and a more sharp surface. Comparison with Other Methods We first show the 3D bone structure in two rendering ways as in Figure 5, where the volume rendering can show the recon- structed surface and the maximum projection can indicate the restored density information. Then we summarize the evaluation metrics in Table 2 to compare with other meth- ods. We can see that Oral-3D has the best performance over other models. Comparing Oral-3D and Auto Encoder with the Residual CNN and GAN, we can see the importance of Figure 6: We show reconstruction results for patients with decoupling the back-projection and deformation process. To wisdom/missing teeth and mark the key features with red be noted, R2N2 achieves the worst performance, where the bounding boxes. We can see that our method can accurately model only learns the shape of the oral cavity but loses de- locate these positions, which can be an important reference tails of teeth. This has indicated the defect when converting during the surgery. the PX image as a collection of multi-view images. Addi- tionally, we see that Auto Encoder has the closest perfor- mance to Oral, although the latter has a more clear surface. networks, i.e. Oral-3D and GAN, we introduce the discrim- This has proved the promotion brought by the adversarial inative network after 100 epochs to alleviate the influence of loss. discrimination loss at the beginning. Identification of Wisdom/Missing Teeth In this paragraph, we show two of the most common cases Results in dental healthcare, e.g., dental implants and tooth pulling, to see if Oral-3D can provide dentist useful reference. Both In this section, we evaluate the reconstruction performance cases require to locate the operation location before the of Oral-3D from different different perspectives. We first surgery. In the first row of Figure 6, three wisdom teeth compare Oral-3D with other methods qualitatively and can be seen clearly on both sides in PX. These features also quantitatively. Then we show the results of special cases for present in the two sides of the reconstruction results. In the some common dental applications. In the end, we do clinical second row, the patient misses two teeth on both sides of trails by evaluating our method on real-word images. the mandible. While the missing place can also be located
Figure 7: We show a workflow to apply Oral-3D to obtain the dental arch curve in real-world applications in this picture. We first take a picture of the patient’s mouth and segment then dental area semi-automatically. Then we use a cubic function to the fit points sampled from the skeletonized image of the binary mask. Table 2: Quantitative Evaluation of 3D Reconstruction Method View Prior D-Net PSNR (dB) SSIM (%) Dice (%) Overall Residual CNN 1 No No 17.46±9.58 72.90±2.09 57.95±7.43 73.54 GAN 1 No Yes 17.71±1.04 69.96±1.91 57.80±7.76 73.78 R2N2 3 No No 18.06±0.94 71.94±1.36 57.71±6.52 73.32 Oral-3D (Auto-Encoder) 1 Yes No 19.04±0.85 76.78±1.65 69.68±4.98 80.56 Oral-3D (GAN) 1 Yes Yes 19.22±0.83 78.27±1.74 71.28±4.69 81.89 Table 3: Evaluation results of different combination of dis- crimination loss (DL), reconstruction loss (RL), and projec- tion loss (PL). DL only DL+PL DL+RL DL+RL+PL PSNR 8.06 18.06(+10.00) 19.14(+11.08) 19.22(+11.16) Figure 8: Although the quality decreases in density details SSIM 46.61 73.02(+26.41) 78.41(+31.80) 78.27(+31.66) for real-word PX, we can still identify each tooth in the re- Dice 35.50 64.53(+29.03) 70.89(+35.39) 71.28(+35.78) construction result. Overall 40.79 75.95(+35.16) 81.66(+40.87) 81.89(+41.10) Table 4: Evaluation results on real-world images is shown in Figure 7. We use cycleGAN (Zhu et al. 2017) to alleviate the colour variance between the training and test- Dataset PSNR SSIM Dice ing PX images. As shown in Table 4, the drop mainly comes from the PSNR and SSIM, which is because the colour vari- Real 17.36±0.70 69.30±2.03 71.44±3.66 ance in different CBCT machines. From Figure 8 we can Synthesized 19.22±0.83 78.27±1.74 71.28±4.69 that although the quality decreases in density details, we can still identify each tooth in the reconstruction result. accurately in the reconstruction image. Conclusion Ablation Study In this paper, we propose a two-stage framework to recon- struct the 3D structure of the oral cavity from a single 2D To reveal the factors that influence the reconstruction quality PX image, where individual shape information of the dental of the generation network, we also do an ablation by chang- arch is provided as prior knowledge. We first utilize a gen- ing the combination of the loss functions. As shown in Ta- erative model to back-project the 2D image into 3D space, ble 3, we see that the model shows the worst performance if then deform the generated 3D image into a curved plane to trained only with the adversarial loss. This is mainly because restore the oral shape. We first use synthesized data to com- the adversarial loss can not bring voxel-wise optimization. pare with different methods, then evaluate the model with We can also see that the major improvement comes from the real-world data to see the feasibility in clinical applications. reconstruction loss, while the projection loss brings much Experimental results show that our model can recover both less promotion, especially when trained with the reconstruc- the shape and the density information in high resolution. We tion loss together. This is also reasonable as the reconstruc- hope this work can help improve dental healthcare from a tion loss can supervise the generation network to learn more novel attitude. detailed information. Clinical Trials Acknowledgement In the end, we evaluate Oral-3D on real-world data from 6 We thank Dr. Liang Chengwen, Dr. Wangbin, and Dr. Wang patients. The workflow of collecting dental arch information Weiqian for collecting data.
References Mao, X.; Li, Q.; Xie, H.; Lau, R. Y.; Wang, Z.; and Braun, S.; Hnat, W. P.; Fender, D. E.; and Legan, H. L. 1998. Paul Smolley, S. 2017. Least squares generative adversarial The form of the human dental arch. The Angle Orthodontist networks. In Proceedings of the IEEE International Confer- 68(1): 29–36. ence on Computer Vision, 2794–2802. Brooks, S. L. 2009. CBCT dosimetry: orthodontic consid- Mirza, M.; and Osindero, S. 2014. Conditional generative erations. In Seminars in Orthodontics, volume 15, 14–18. adversarial nets. arXiv preprint arXiv:1411.1784 . Elsevier. Momin, M. A.; Okochi, K.; Watanabe, H.; Imaizumi, A.; Omura, K.; Amagasa, T.; Okada, N.; Ohbayashi, N.; and Choi, H.; and Lee, D. S. 2018. Generation of structural MR Kurabayashi, T. 2009. Diagnostic accuracy of cone-beam images from amyloid PET: application to MR-less quantifi- CT in the assessment of mandibular invasion of lower gin- cation. Journal of Nuclear Medicine 59(7): 1111–1117. gival carcinoma: comparison with conventional panoramic Choy, C. B.; Xu, D.; Gwak, J.; Chen, K.; and Savarese, S. radiography. European journal of radiology 72(1): 75–81. 2016. 3d-r2n2: A unified approach for single and multi- Petersen, L. B.; Olsen, K. R.; Christensen, J.; ; and Wen- view 3d object reconstruction. In European conference on zel, A. 2014. Image and surgery-related costs comparing computer vision, 628–644. Springer. cone beam CT and panoramic imaging before removal of Costa, P.; Galdran, A.; Meyer, M. I.; Abràmoff, M. D.; impacted mandibular third molars. Dentomaxillofacial Ra- Niemeijer, M.; Mendonça, A. M.; and Campilho, A. 2017. diology 43(6): 20140001. Towards adversarial retinal image synthesis. arXiv preprint Prajapati, S. A.; Nagaraj, R.; and Mitra, S. 2017. Classi- arXiv:1701.08974 . fication of dental diseases using CNN and transfer learning. Cui, Z.; Li, C.; and Wang, W. 2019. ToothNet: automatic In 2017 5th International Symposium on Computational and tooth instance segmentation and identification from cone Business Intelligence (ISCBI), 70–74. IEEE. beam CT images. In Proceedings of the IEEE Conference Wang, Z.; Bovik, A. C.; Sheikh, H. R.; and Simoncelli, E. P. on Computer Vision and Pattern Recognition, 6368–6377. 2004. Image quality assessment: from error visibility to Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; structural similarity. IEEE transactions on image process- Warde-Farley, D.; Ozair, S.; Courville, A.; and Bengio, Y. ing 13(4): 600–612. 2014. Generative adversarial nets. In Advances in neural Wu, J.; Zhang, C.; Zhang, X.; Zhang, Z.; Freeman, W. T.; information processing systems, 2672–2680. and Tenenbaum, J. B. 2018. Learning shape priors for Henzler, P.; Rasche, V.; Ropinski, T.; and Ritschel, T. 2018. single-view 3d completion and reconstruction. In Pro- Single-image Tomography: 3D Volumes from 2D Cranial X- ceedings of the European Conference on Computer Vision Rays. In Computer Graphics Forum, volume 37, 377–388. (ECCV), 646–662. Wiley Online Library. Yang, G.; Cui, Y.; Belongie, S.; and Hariharan, B. 2018. Huang, G.; Liu, Z.; Van Der Maaten, L.; and Weinberger, Learning single-view 3d reconstruction with limited pose K. Q. 2017. Densely connected convolutional networks. In supervision. In Proceedings of the European Conference on Proceedings of the IEEE conference on computer vision and Computer Vision (ECCV), 86–101. pattern recognition, 4700–4708. Ying, X.; Guo, H.; Ma, K.; Wu, J.; Weng, Z.; and Zheng, Y. 2019. X2CT-GAN: reconstructing CT from biplanar X-rays Imangaliyev, S.; van der Veen, M. H.; Volgenant, C. M.; Kei- with generative adversarial networks. In Proceedings of the jser, B. J.; Crielaard, W.; and Levin, E. 2016. Deep learn- IEEE conference on computer vision and pattern recogni- ing for classification of dental plaque images. In Interna- tion, 10619–10628. tional Workshop on Machine Learning, Optimization, and Big Data, 407–410. Springer. Yun, Z.; Yang, S.; Huang, E.; Zhao, L.; Yang, W.; and Feng, Q. 2019. Automatic reconstruction method for high-contrast Isola, P.; Zhu, J.-Y.; Zhou, T.; and Efros, A. A. 2017. Image- panoramic image from dental cone-beam CT data. Com- to-image translation with conditional adversarial networks. puter methods and programs in biomedicine 175: 205–214. In Proceedings of the IEEE conference on computer vision and pattern recognition, 1125–1134. Zhu, J.-Y.; Park, T.; Isola, P.; and Efros, A. A. 2017. Un- paired image-to-image translation using cycle-consistent ad- Kingma, D. P.; and Ba, J. 2014. Adam: A method for versarial networks. In Proceedings of the IEEE international stochastic optimization. arXiv preprint arXiv:1412.6980 . conference on computer vision, 2223–2232. Kolev, K.; Brox, T.; and Cremers, D. 2012. Fast joint esti- mation of silhouettes and dense 3d geometry from multiple images. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence 34(3): 493–505. Lee, J.-H.; Kim, D.-H.; Jeong, S.-N.; and Choi, S.-H. 2018. Detection and diagnosis of dental caries using a deep learning-based convolutional neural network algo- rithm. Journal of dentistry 77: 106–111.
You can also read